RESEARCH DOSSIER
Deep Representation Learning in Insurance Analytics
Abstract. Deep representation learning converts heterogeneous insurance information into learned numerical features that can support pricing, underwriting, reserving, claims analysis, and risk monitoring. Its scientific significance lies in the relation between the information available at a legitimate decision time, the objective that shapes a representation, and the actuarial quantity subsequently estimated. A learned embedding is a function of observations, training data, and optimization constraints. Its apparent geometry alone does not establish a causal risk factor, an economically meaningful characteristic, or a stable basis for contractual decisions. This monograph develops a comprehensive analysis of representation objectives, identification, tabular and high-cardinality variables, temporal exposure histories, language, images, multimodal fusion, insurance-specific statistical targets, transfer, selection bias, robustness, uncertainty, fairness, and governance. It distinguishes documented research results from mathematical deductions and proposed validation procedures, and evaluates original actuarial evidence concerning categorical embeddings, street-view information, combined actuarial neural networks, and telematics data integration. Formal derivations connect frequency and severity likelihoods to embeddings, expose a simple source of selection-induced bias, and demonstrate why prediction invariance under latent coordinate transformations precludes a naïve interpretation of individual embedding dimensions. A worked numerical illustration isolates the effect of changes in estimated frequency on expected loss without claiming an empirical gain. The principal conclusion is that the value of a representation is conditional on its admissible observation history, target definition, incremental performance over strong conventional baselines, calibration, stability, and verifiable institutional use. Representation learning can extend the measurable insurance information set while leaving contractual eligibility, exposure measurement, and scientific inference as distinct responsibilities. This work offers original analytical synthesis and evaluation criteria without claiming a newly invented learning theory, actuarial doctrine, or externally validated dataset result.
Keywords: Deep representation learning; insurance analytics; actuarial science; embeddings; latent features; high-cardinality categorical variables; neural networks; self-supervised learning; contrastive learning; tabular data; telematics; insurance documents; images; multimodal fusion; claim frequency; claim severity; exposure offsets; calibration; dataset shift; data leakage; fairness; model governance; uncertainty; underwriting; pricing; reproducibility.
Introduction
An insurance decision occurs at a particular time, over a defined contractual exposure, with observable information that has been collected through specific institutional processes. Risk prediction connects this information to uncertain outcomes. Modern insurers possess sources of information far richer than the variables in a conventional rating table: claim narratives, policy wording, photographs of insured property, categorical product histories, telematics streams, inspection reports, call-center records, and repeated transactions. These sources vary in structure, measurement error, temporal availability, consent, confidentiality, and legal admissibility. Their sheer volume does not make them risk measures. The first scientific problem is to construct a representation that can be estimated, tested, and used in an actuarially defined prediction problem.
Representation learning addresses this problem by estimating an encoder that maps observed information into features useful under some training criterion. The representation-learning review by Bengio, Courville, and Vincent (2013) sets out a general program of learning features that reveal or separate explanatory variation. Insurance adds a demanding requirement: the prediction target must be linked to covered events, exposure units, coverage terms, observation periods, and claim development. A vector that predicts whether a document contains the word “accident” can be effective for routing documents while offering no warranted information about future loss costs. Conversely, a vector that summarizes the pre-coverage history of driving behavior may be relevant to prospective claim frequency even if its dimensions have no intuitive actuarial names.
The scientific object here is the learned representation in its complete predictive setting. This includes the observation process and timestamp, the mapping from raw records to embeddings, the objective used for training, the prediction or decision to which the embedding contributes, and the validation design under which usefulness is judged. The unit of inquiry is therefore a representation–target–population relationship. A learned feature cannot be evaluated independently of the task for which it is proposed. The distinction matters whenever an image encoder trained to distinguish surfaces, an unsupervised text model trained to reconstruct language, or a tabular encoder trained to predict missing entries is subsequently used to estimate insured losses.
Empirical actuarial research supplies concrete examples. Blier-Wong, Lamontagne, and Marceau (2024) integrate street-view image embeddings into generalized linear insurance models and report information useful for predicting claim frequency for certain perils; their insurance-pricing image study also addresses social and technical difficulties. Avanzi, Taylor, Wang, and Wong (2024) examine high-cardinality actuarial categories and propose a network incorporating generalized linear mixed-model structure; their high-cardinality actuarial comparison demonstrates why categorical representation and random-effect inference must be considered together. Duval, Boucher, and Pigeon (2024) test combined actuarial neural networks for cross-sectional and longitudinal automobile claim counts; their telematics claim-count models provide an insurance-specific comparison rather than an endorsement of every neural architecture.
These findings invite a synthesis that is both statistical and institutional. Representation learning can create new dimensions of observable variation, change the effective granularity of risk pools, and reduce the handcrafting of predictors. It can also import claims labels into predictors, encode geographic or demographic proxies, amplify policy-selection effects, or mistake a change in data collection for a change in underlying hazard. Evaluating a representation requires separating its predictive gain from its provenance, reliability, and permissible use.
The argument proceeds from the formal meaning of representation through training objectives, modality-specific applications, actuarial estimation, evaluation and governance. Each published study is interpreted within the limits of its data and method. Algebraic conclusions are derived explicitly. No empirical performance figures are invented for a model or insurance portfolio that has not been studied. The final result is a set of defensible conclusions about how to test and use learned representations in insurance analysis.
Research Object, Scope, and Evidentiary Method
The research object is an estimated encoder and its relation to a defined insurance outcome. Let an observation for unit i at decision time t be X_i(t), with contractual terms C_i(t), exposure E_i(t), and future target Y_i(t,t+h). The encoder f_θ maps the data that are actually admissible and available at t to Z_i(t)=f_θ(X_i^{≤t},C_i^{≤t}); a downstream predictor g_β uses Z, conventional factors and exposure information. The temporal superscript matters. A representation that consumes a record created after the prediction date is solving another problem, regardless of its apparent predictive accuracy.
A prospective motor rating problem may use the policy record, vehicle information, and consented driving history through the quotation date to predict claims during an ensuing period. A claims triage problem instead uses the notice of loss and evidence available at first notification to predict handling complexity or probable settlement cost. A reserve evaluation uses reported claims and payments known at the valuation date to estimate ultimate amounts. Each application has a distinct target, population, exposure scale, and legal or operational consequence. Sharing an encoder across tasks is an empirical hypothesis to test, not an assumption that identical representations are sufficient.
The evidence base combines foundational machine-learning papers, original peer-reviewed actuarial articles, and primary institutional texts. Research sources are differentiated by genre. The seminal CANN editorial by Wüthrich and Merz (2019) develops an illustrative methodological proposition; it does not itself constitute a broad empirical trial. The studies by Blier-Wong et al. (2024), Avanzi et al. (2024), Duval et al. (2024), and Peiris et al. (2024) report particular actuarial methods and evaluations; their observations do not establish universal out-of-portfolio performance. Institutional publications are authoritative for their own stated regulatory or policy positions, within the jurisdiction and legal status they specify.
The method uses four layers of reasoning. First, formal analysis defines the representation, learning objective, output distribution, and information set. Second, comparative evidence examines what peer-reviewed papers actually demonstrate. Third, deduction derives consequences for insurance-specific problems such as exposure offsets, selection of observable telematics, and latent nonidentifiability. Fourth, editorial validation tests whether a source supports each externally derived claim, whether its limitations are disclosed, and whether a proposed operating criterion can be checked with real data. A proposed QA procedure in this monograph is an analytical recommendation, not evidence that a given insurer has implemented it.
The limits of the investigation are equally explicit. No individual-level proprietary policy or claim dataset has been supplied. This study presents no newly fitted network, external prospective field trial, or validated performance improvement measured in an undisclosed portfolio. It does not make jurisdiction-wide assertions where rules are national or subnational. Its numerical illustration uses stated hypothetical assumptions solely to demonstrate arithmetic and evaluation logic. Literature availability varies between open full-text journal articles, publisher abstracts, official technical records, and statutory portals. These evidentiary differences are recorded in the associated source register.
What a Learned Representation Is
A representation is a function of observable data. Write Z=f_θ(X), where X may be a mixed numerical and categorical vector, a sequence of sensor events, a document, an image, or a collection of modalities. The parameter θ is learned from observations through a defined objective. A downstream mapping g_β(Z,C,E) produces an estimate of a contractual target. The encoder may be fitted separately through self-supervision, simultaneously with the target model, or using a combination of objectives. In each case Z is associated with the distribution and preprocessing on which θ was estimated.
The statistical meaning of Z can be described in three ways. As compression, it discards some details of X while preserving others. As prediction-oriented transformation, it makes certain responses easier to forecast. As a learned coordinate system, it organizes variation according to the chosen objective and architectural constraints. These descriptions are complementary but none establishes that a particular coordinate represents a uniquely recoverable physical cause. A category vector for occupation, for example, can identify predictive similarities across occupation codes while reflecting uneven exposure, reporting, and class definitions. A document vector can reflect the linguistic style of the claims adjuster as well as the event described.
The elementary sufficiency ideal would require that, for a chosen insurance target Y and controlling covariates C, the conditional distribution satisfy P(Y|X,C)=P(Y|Z,C). That equality is demanding and almost never proved for real-world encoders. It is more reasonable to ask whether adding Z reduces estimated prospective risk under a proper scoring rule, improves calibration within relevant segments, or supports an operating decision in independent data. The expectation of sufficiency is a design aim; observed predictive gain and robustness are evidence.
The dimension d of Z is an engineering choice, and it does not measure the number of meaningful causes. A small embedding may underfit important sparse effects; a very large embedding may memorize policy or territorial identifiers. Regularization, data volume, inductive biases, input treatment, and fine-tuning all affect this balance. Bengio et al. (2013) identify the significance of learned feature structure, while Locatello and colleagues' identifiability study (2019) demonstrates the limitations of discovering uniquely disentangled factors from unlabeled observations without additional assumptions. In insurance, substantive interpretation requires independent evidence beyond attractive plots of embedding space.
Representations are also relational. If the downstream objective is the expected number of covered claims, an encoder for photographs may emphasize building material, roof condition, or neighboring vegetation, but it may also emphasize photography season or locality. The insurer must test which of those aspects are stable, legally relevant, and acquired on an acceptable basis. The phrase “risk representation” is scientifically earned through the subsequent predictive and institutional tests, not by naming the embedding.
Representation Learning Objectives
Training objectives select the information an encoder tends to preserve. In supervised learning, parameters minimize an empirical risk of the form L_sup(θ,β)=Σ_i w_i ℓ(Y_i,g_β(f_θ(X_i),C_i,E_i)), with loss ℓ appropriate to the target and weights w_i justified by sampling or exposure. The encoder learns only to the extent the labeled sample contains reliable future outcome information. Claim frequency labels are sparse, delayed, and zero-heavy; claim settlement labels may be censored or revised. These properties distinguish insurance from ordinary balanced classification tasks.
Autoencoding uses a decoder d_φ and reconstruction criterion L_rec=Σ_i ℓ_X(X_i,d_φ(f_θ(X_i))). Reconstruction preserves features useful for reconstructing inputs. Some of these features may be irrelevant to insured loss, and a high-fidelity reconstruction of personal identifiers could increase memorization and privacy risk. Denoising autoencoders introduce corrupted input X̃ and reconstruct the original X; the denoising autoencoder method of Vincent et al. (2008) shows how corruption can encourage robust features in general machine learning. For policies, a corruption mechanism that indiscriminately masks exposure time, coverage limits, or mandatory legally meaningful fields may teach an encoder the wrong invariances. Missingness introduced by training augmentation must be distinguished from real-world missingness generated by applicants, intermediaries, or inspection systems.
Contrastive methods learn to associate positive views of an underlying observation and distinguish other observations. In a simplified normalized setting, with anchor z_i, corresponding positive z_i+, comparison set A_i and temperature τ, the InfoNCE-type loss is −log{exp(sim(z_i,z_i+)/τ)/Σ_{k∈A_i}exp(sim(z_i,z_k)/τ)}. The SimCLR analysis of Chen et al. (2020) demonstrates how augmentations affect visual contrastive learning. Insurance requires a separate argument for every permitted transformation. Crop or brightness variation may preserve roof material but may erase waterline marks, signs of storm damage, or structural defects. Time shifts in driving data can change exposure and behavior. A positive pair should reflect a scientifically justified equivalence relation for the intended decision.
Language modeling, masked-value prediction, sequence forecasting, and cross-modal matching produce further representation objectives. The BERT paper by Devlin et al. (2019) establishes masked-language pretraining and downstream adaptation in natural-language tasks. An insurance corpus introduces domain terminology, abbreviations, negotiated contract clauses, and differences between a claimant's narrative and an insurer's conclusion. Pretraining on such material can improve textual adaptation, but correctness in coverage or causation analysis must be checked against labeled, temporally valid and legally reviewed tasks.
Multitask learning minimizes a weighted sum of losses, for example L=λ_f L_frequency+λ_s L_severity+λ_r L_reconstruction. The weights are not a decorative hyperparameter: they determine which information dominates the encoder. A target with millions of routine examples can overwhelm a rarer catastrophic outcome. Conversely, an attempt to emphasize scarce extremes may cause unstable gradients or noisy representations. For each proposed objective, the researcher must state its training labels, observation times, class or exposure weights, invariances, and relation to a business-relevant estimand.
The strongest scientific inference from this comparison is conditional. No generic objective guarantees an insurance-valid embedding. An objective is justified when its information-preserving tendencies can be stated, tested against target-relevant alternatives, and shown to generalize under the intended policy and claim population.
Embedding Structures and Statistical Identifiability
Categorical embeddings assign each category c a row e_c from a learned table E∈R^{K×d}. Sequence encoders map an ordered history into a state or pooled vector. Image encoders transform pixel arrays into feature maps or pooled representations. Text encoders map tokenized documents into context-dependent representations. The resulting vectors occupy numerical spaces shaped by feature preprocessing, parameterization, sampling and the loss function. Distances or clusters are interpretable only relative to those choices.
A simple mathematical observation establishes the coordinate problem. Consider a prediction g_β(z)=βᵀz, where z=f_θ(x). For any invertible matrix A, define z′=Az and β′=A^{−T}β. Then β′ᵀz′=βᵀA^{−1}Az=βᵀz. Consequently, the same prediction may arise from different latent coordinate systems. A rotation can change the numerical value and apparent interpretation of each coordinate while preserving every fitted score. This is a direct algebraic result, not a newly proposed representation theory. It shows that individual embedding-coordinate narratives require a separately fixed basis or external validation.
Even stronger concerns apply to nonlinear encoders. Several functions can yield nearly identical predictions under the observed data distribution while behaving differently under new policies, geographies, or claims conventions. A semantically appealing embedding diagram cannot resolve that ambiguity. The identifiability limitations emphasized by Locatello et al. (2019) reinforce a general caution about unsupervised factor recovery, although the specific impossibility result should not be generalized to every supervised or structurally identified insurance application.
For actuarial analysis, interpretation should attach to stable groups of observables and prediction consequences. One can perturb a roof characteristic while holding suitable comparable attributes fixed, assess whether the prediction changes consistently, and investigate whether the source of the characteristic is reliable. One can test neighborhood or occupation embeddings under held-out jurisdictions. One can aggregate attribution to the original input fields while reporting uncertainty and correlations. These methods yield evidence about fitted behavior; they do not automatically establish effects of interventions on risk.
Embedding geometry must also be treated as time-indexed. A category identifier for an occupation class that undergoes recoding in a later underwriting year is a different observation regime. An insurer may need an explicit versioned mapping instead of comparing the new learned vector to an old one without alignment. Category labels can remain textually constant while their insured populations change. The definition of a representation therefore includes the vocabulary, preprocessing schema, training cutoff and production version.
High-Cardinality and Structured Insurance Records
Insurance portfolios contain categorical variables with potentially thousands of levels: vehicle models, occupations, postcodes, commercial activity codes, repair facilities, broker identifiers, product combinations, and claim-cause classifications. One-hot encoding expands each level into a separate dimension. For rare levels it can estimate highly variable effects, while arbitrary category grouping can discard genuine heterogeneity. Learned entity embeddings offer partial pooling through shared downstream parameters, with predictive information carried in fewer dimensions than a full set of indicators. The quality of that pooling depends on sample size, exposure weighting, regularization, and the construction of the category itself.
The GLMMNet study by Avanzi et al. (2024) is central here. Its combination of neural features and mixed-model structure demonstrates a useful distinction between a predictive embedding and a random effect with an explicit uncertainty interpretation. In traditional generalized linear mixed modeling, category effects are partially pooled through a distributional assumption; the fitted level effects and their uncertainty are connected to that model. A neural embedding can be powerful for prediction, but its axes cannot simply be read as equivalent random-effect estimates. Treating the two as interchangeable would import inferential properties that the embedding has not established.
An insurance-specific design retains several forms of structure. Exposure is represented as an offset or explicit denominator, rather than left for the network to rediscover. Coverage period, deductible, limit, line of business and claim definition are checked as semantic inputs. Numeric continuous variables retain a calibrated scale and explicit missingness treatment. High-cardinality fields are versioned to account for newly appearing levels. Administrative identifiers such as claim number, policy number, individual customer ID and internal routing code are excluded unless their use is independently justified, because they can encourage memorization or encode a future outcome through operational workflow.
Tabular deep learning remains an empirical contest. The TabNet study by Arik and Pfister (2021) proposes a sequential-attention model and self-supervised treatment of tabular features. The benchmark by Gorishniy et al. (2021) identifies strong tabular neural baselines and finds no universally superior approach. The comparison by Grinsztajn et al. (2022) finds that tree-based methods remain strong across many medium-sized tabular datasets. Those findings concern their particular benchmarks, not a universal actuarial leaderboard; nevertheless they establish the need for GLM, GAM, mixed-model and tree-ensemble comparators in insurance experiments.
The relevant question is whether embedding-based learning improves a contractual target after the additional complexity, drift sensitivity and governance cost are counted. If an occupation embedding is competitive with a simpler regularized categorical regression, the scientific result may be equivalence at current data scale. If it improves out-of-time log-likelihood and calibration while maintaining consistent category behavior, the added representational capacity has demonstrated value. Neither outcome should be prejudged by the architecture's novelty.
Exposure Histories, Telematics, and Longitudinal Learning
Insurance records often include repeated measurements of the same insured, vehicle, structure, firm, or contract. A history encoder h_θ(x_{i1},...,x_{it}) may summarize past events, usage, behavior, claims notifications, or financial indicators. Its scientific usefulness depends on the sequence boundary. Information recorded during a claim investigation cannot legitimately enter a model predicting whether that claim would occur beforehand. Likewise, a future change in coverage is not a historical feature at the current quotation date. For repeated annual contracts, leakage may arise even when every row is individually stamped correctly: the same insured appears in training after a later outcome and in testing before it.
Telematics illustrates the opportunity and the problem. Acceleration, braking, mileage, time of day, and driving patterns can be summarized by manually engineered statistics or learned from sequences and speed–acceleration representations. The Duval–Boucher–Pigeon study (2024) develops combined actuarial neural models for cross-sectional and longitudinal claim counts and reports favorable performance against selected log-linear models with manually engineered telematics features. The result is specific to its study designs and does not establish that learned telematics features are an unbiased measure of individual driver disposition.
The denominator is central. A driving-behavior summary observed over 100 miles and one observed over 20,000 miles have different measurement precision. An annual claim count may depend on the duration or volume of exposure, which is partly determined by customer behavior. A sequence model should distinguish intensity of risky behavior, volume of opportunity, and observation length. In a frequency model with claim count N_i and earned exposure e_i, it is coherent to use log e_i as an offset while providing measured mileage as a predictor only when the exposure definition and potential dependence are made clear. Replacing an exposure offset with an embedding can obscure a basic actuarial accounting identity.
Longitudinal information also creates dependence. Policies renewed by the same driver share unobserved characteristics, and incidents are influenced by time-varying road and usage conditions. Random train/test row splitting can place near-duplicate histories of a person into both sets. A prospective test should split by decision time and, where the deployment requires independence across insureds or groups, by insured or fleet. Temporal dependence should be reflected in uncertainty calculations, and claim reporting delays require a maturity strategy. Sequence encoders can process irregular time intervals, but interval encodings and missing-observation indicators must be explicit and tested.
Another problem is who generates a telematics record. Participation in usage-based products may be voluntary, offered to selected customers, or influenced by privacy preferences. The data-integration study by Peiris et al. (2024) addresses uneven telematics availability and favorable selection using calibration weighting; it evaluates simulations and synthetic-data applications. That evidence supports a selection-aware methodology, while its synthetic empirical component cannot itself establish market-wide field performance. A sophisticated temporal encoder fitted only to volunteers can still transfer poorly to the full book of business.
The appropriate representation of driving behavior is therefore a temporally bounded, exposure-aware function estimated on a population with documented participation and missingness. Its performance should be assessed separately among participants, new customers, renewal cohorts, regions, and mileage strata. Changes in device firmware or data delivery can modify measured events even if the underlying risk remains unchanged. In such a case, the representation has experienced a measurement shift before any actuarial hazard change has been demonstrated.
Text, Images, and Documents as Insurance Information
Natural-language material enters insurance through applications, inspection narratives, policy clauses, adjuster notes, claimant statements, repair quotations and litigation records. A text encoder can summarize thematic content or retrieve relevant passages, but these documents are produced by agents with distinct incentives and at distinct moments in an insurance process. A claimant's narrative is evidence of a report, an adjuster's note is an operational interpretation, and the contract wording determines legally relevant obligations in the applicable jurisdiction. A single general-purpose text embedding does not merge their evidentiary status.
Pretrained contextual language representations, exemplified by the BERT architecture of Devlin et al. (2019), are a foundation for downstream document modeling. For insurance, fine-tuning labels should reflect an accurately defined task: sorting incoming claims, classifying document type, extracting a policy limit, identifying an exclusion clause, or estimating a later operational outcome. Training labels based on a disputed adjudication or an investigation can convey annotation uncertainty. A useful experiment measures inter-annotator agreement, stratifies by document source and policy year, and checks whether the encoder relies on boilerplate that changed in a new form rather than on the legally relevant content.
A text representation used for future loss prediction raises a temporal distinction. Prior underwriting notes can be legitimate antecedent information in an authorized setting. Notes written after a serious claim, even if found in the same enterprise data warehouse, cannot become predictors of that claim's prior probability. A fraud model may deliberately use post-notification evidence, but its target then concerns a later investigative or adjudicative status. Failing to define the decision time turns document embeddings into a powerful channel for post-outcome leakage.
Images of insured assets introduce similar complexities. An image encoder may record visual properties of buildings, vehicles or damage. The Blier-Wong–Lamontagne–Marceau insurance study (2024) shows how low-dimensional street-view representations can enter a generalized linear model and add predictive information for certain property claim frequencies. The authors also discuss technical and societal difficulties. Street-view data can be outdated, unavailable for rural properties, influenced by camera coverage, or correlated with neighborhood wealth and infrastructure. Those characteristics matter to external validity and to whether the learned signal can support lawful and fair classification.
Consider two photographs of the same property taken in different seasons. Their learned representations may differ despite constant construction materials. Conversely, a reconstruction process may collapse flood-damage marks and harmless lighting variation into similar features. The research question is which image variations are related to the target and which should be treated as nuisance. In self-supervised work, declared augmentations determine learned invariances; the contrastive-learning findings of Chen et al. (2020) make the dependence on augmentation design explicit. Insurance requires domain scrutiny of those transformations rather than importing a standard computer-vision pipeline without examination.
Digitized documents introduce another modality: a scanned invoice or policy endorsement combines text, layout and image information. A layout-sensitive representation may be useful for field extraction and claims workflow, yet document authenticity, version history, and contractual authority need separate verification. A model that classifies similar-looking forms cannot substitute for checking whether the signature, endorsement, or policy schedule was in force. A scientifically strong pipeline preserves links from embeddings back to the input document, its date, its provenance and any human correction.
Multimodal Integration with Contract and Exposure Data
Let X^T denote tabular contract features, X^S time-series information, X^D documents, and X^I images. A multimodal system estimates z_T=f_T(X^T), z_S=f_S(X^S), z_D=f_D(X^D) and z_I=f_I(X^I), then maps the available representations and contractual terms to a distribution of Y. Early fusion concatenates raw or intermediate features; late fusion combines predictions; intermediate fusion learns cross-modal relationships. These are implementation families with different assumptions about sample alignment, data volume and missing modalities. Scientific evaluation should compare them under the same decision dates and labels.
An insurer typically has complete core contract records but sparse or selective image, text or telematics data. A naïve multimodal system trained only on policies with all modalities changes the sampling population. One approach includes missingness indicators and modality-specific branches, but this does not by itself correct nonrandom availability. Another approach uses probabilistic or weighted estimation under an explicit observation mechanism. The Peiris et al. (2024) integration framework offers direct actuarial evidence of the importance of selection and weighting when combining conventional and telematics sources.
Availability can depend on region, policy value, distribution channel, consumer participation, device type and anticipated risk. Let M_i denote the presence of an optional modality. If P(M_i=1|X_i,Y_i) varies with the outcome even after conditioning on observed variables, complete-case training need not estimate the distribution for the full portfolio. A missingness flag can help a predictive model exploit patterns of collection, but it cannot independently recover the missing outcome relationship. Governance additionally asks whether deriving a rating factor from a person's decision to withhold optional data is lawful and fair.
Cross-modal alignment is a data engineering and statistical issue before it is an architectural one. The photographed building must be the insured building; the image date must precede the prediction date; the telematics device must refer to the correct vehicle and policy period; the document must have a known contract and endorsement status. A single incorrect property join can generate a high-dimensional yet meaningless signal. Quality controls should include entity-resolution confidence, timestamps, modality provenance, deterministic as-of joins and an auditable record of discarded matches.
For comparison, a robust baseline uses only structured features from the same population. A second baseline adds manually constructed features from the new modality; a third uses pretrained frozen embeddings; a fourth fine-tunes or jointly trains the encoder. Each comparison asks a distinct question: whether the modality adds information, whether automatic representation beats expert extraction, whether fine-tuning improves a generic model, and whether additional complexity is justified by the gain. Comparisons made on different complete-case populations cannot isolate the value of the encoder.
The resulting multimodal representation is a collection of estimated measurements, not an all-purpose factual account of the risk. Its scientific scope is limited by data accessibility, temporal alignment, training target, missingness process and the predictive population in which it has been validated. Integrating more modalities can increase information and simultaneously increase the number of ways to introduce selection, consent and provenance errors.
Frequency, Severity, and Reserving Targets
A learned representation becomes actuarially meaningful when connected to a well-defined insurance estimand. For claim frequency, let N_i be claims over exposure e_i>0, and let z_i=f_θ(x_i). Under a Poisson working model, N_i|x_i,e_i follows Poisson(μ_i), with log μ_i=log e_i+α+βᵀz_i+γᵀc_i. The log e_i term is an exposure offset. Its coefficient is fixed at one because it expresses the linear scaling assumption of the frequency model. Training can minimize the negative Poisson log-likelihood Σ_i[μ_i−N_i log μ_i+log(N_i!)]. If observed variability substantially exceeds the conditional Poisson variance, an overdispersed count model, such as a negative binomial, may be more appropriate. Model comparison must examine deviance, calibration and stability, not only overall mean absolute error.
For individual claim severity, let S_ij>0 denote a positive payment or ultimate loss for claim j on contract i. A conditional distribution may be gamma, lognormal, another positive-valued family, or a flexible conditional model. The selected likelihood should correspond to the chosen censoring, limits, deductibles, inflation treatment and claim-development status. Zero payments, reopened claims and large losses need explicit treatment rather than implicit deletion. If a model predicts positive severity conditional on claim occurrence, it cannot be interpreted without the frequency and coverage conventions that define that condition.
For a simplified compound frequency-severity model, expected aggregate loss per contract is E[L_i|x_i]=E[N_i|x_i]·E[S_i|x_i] when the assumed conditional relationship justifies the factorization and claim marks are modeled consistently. A shared encoder can feed both components, but this may help or harm: the representation best suited to common small claims can miss rare high-severity mechanisms. Separate encoders, shared lower layers and multitask heads should be compared with the same held-out claim development and tail-sensitive metrics. Correlation between claim frequency and severity, policy limits and aggregate events may require a richer joint distribution.
Combined actuarial neural networks offer one route for integrating learned nonlinearity while retaining an explicit classical predictor. Wüthrich and Merz's Yes, We CANN! editorial (2019) presents the underlying concept; Duval et al. (2024) extend it to telematics count models. In a multiplicative version one may write log μ_i=log e_i+η_GLM(c_i)+r_θ(x_i), where r_θ is a learned correction. This parametrization preserves the known offset and exposes the baseline actuarial structure, but it does not automatically make the correction causally interpretable.
Loss reserving creates a different target. At valuation date v, an outstanding claim's ultimate amount must be predicted with only payments, reports and development information observed by v. Encoders may summarize claim narratives and payment sequences, but data are right-truncated by the valuation cutoff and affected by reporting and settlement mechanisms. A representation trained on fully settled historical claims can misrepresent open claims at early development stages if its training sample or labels are not adjusted for development and selection. The target may be ultimate indemnity, future incremental payments, a payment schedule, or an aggregate reserve; these are distinct distributions and error functions.
The relation between representations and insurance accounting is therefore explicit: an encoder provides predictors, while the actuarial model defines what is being estimated and how uncertainty and legal coverage enter the outcome. Without that target discipline, a high-performing neural network may solve a convenient data-label problem with no valid premium or reserving interpretation.
Transferable Risk Features and Dataset Shift
An encoder pretrained on a large external dataset can be transferred to a smaller insurance task. For text, a general-language model may be adapted to claims descriptions. For images, a pretrained visual encoder may be used to form property features. For structured data, an encoder trained on one portfolio can be transferred to another line or territory. Transfer reduces the need to estimate every feature from scarce target labels, but also imports source-domain measurement and selection assumptions. Its usefulness is a question about the source and target joint distributions.
Let P_s(X,Y) and P_t(X,Y) denote source and target populations. A predictor evaluated under P_s may perform differently under P_t if either the distribution of features or the conditional outcome relation changes. Under pure covariate shift, P_s(Y|X)=P_t(Y|X) while P_s(X) differs, but even that equality can be fragile in insurance when coverage, claims definitions or loss prevention practices change. The Dataset Shift (2009) provides the broader mathematical vocabulary for these distributional changes. An encoder's apparent stability in latent space is not proof that the target distribution has remained stable.
The WILDS benchmark by Koh et al. (2021) demonstrates important out-of-distribution losses across several noninsurance domains. It supports the general need to test deployment shifts, while its datasets do not provide measured insurance transfer gains. In insurance, the relevant partitions include calendar period, geographic region, insurer or product, distribution channel, catastrophe regime, and customer tenure. A model may transfer across insurers only after contractual definitions, reporting conventions and coverage units are reconciled. Without this harmonization, equal feature labels do not indicate equal targets.
Transfer evaluation should distinguish frozen-feature linear probing, limited fine-tuning, full fine-tuning, and task-specific training from scratch. These comparisons control for labeled-data budgets and tuning effort. When a pretrained encoder performs well with few labels, the researcher should ask whether it has learned general visual or linguistic factors, whether its pretraining overlaps test cases, and whether the insurance output is calibrated after adaptation. Preventing hidden overlap in pretraining datasets may require documentation from external model suppliers.
A shifted embedding can also be deceptive. Drift of the marginal representation P(Z) may reflect seasonal composition, adoption of a new device, or recoding of categories; unchanged P(Z) does not guarantee unchanged P(Y|Z). Monitoring both input quality and outcome-linked calibration is necessary. In periods when ultimate claims have not matured, short-term proxy metrics can support alarms but cannot verify final loss performance. A formal feedback schedule must reflect claim development rather than merely the pace of model scoring.
Sampling, Missingness, Censoring, and Leakage
Representation learning inherits the observational design of the insurance data from which it learns. A policy administration database usually contains insured risks accepted by a particular underwriting system, subject to its product rules, geographic availability, price levels and customer decisions. A claims database further conditions on claims being reported, captured and adjudicated; an inspection-image archive conditions on properties being inspected; telematics data condition on installation, enrollment, compliance and functioning hardware. Each condition changes the distribution of observed cases. The resulting dataset is not necessarily a random sample of an addressable population, and a successful embedding can encode selection patterns as readily as hazard characteristics.
For a target population T and a selected sample S, the question is whether E_T[L|Z=z] can be estimated from observations in S. In general, E_S[L|Z=z] equals the target conditional expectation only under specified selection and transport assumptions. An analyst who learns embeddings solely from inspected high-value buildings and subsequently deploys them to all residential properties must justify both the missing-image mechanism and the transport of loss relations. Reweighting can sometimes correct measured selection probabilities, but it requires overlap and an adequate model of selection. For groups effectively absent from the development data, neither representation learning nor propensity weighting creates empirical information.
Missing inputs need modality-specific handling. A missing telematics stream may represent nonenrollment, disconnection, device failure, insurance cancellation or a privacy choice; these states are substantively different. Missingness indicators can be predictive, and may also encode administrative decisions or prohibited proxies. A text field that is empty because a claim is new differs from a text field missing after migration between claims systems. Standard masking tokens are computationally convenient only after the observation process is analyzed. Sensitivity analysis should compare complete-case, masked, imputed and missingness-aware models on the same target cohort.
For claims and reserves, censoring and truncation change the information structure. Claim payments observable through a valuation date may be incomplete; claim notices arrive with reporting delays; litigation and reopened files produce long development tails. Denote by H_i(v) the admissible history observable at valuation time v. Then the admissible embedding is Z_i(v)=f_θ(H_i(v)), and the target might be future payments L_i(v,∞). Every token and every fitted preprocessing statistic must be obtainable at v. A representation that includes subsequent settlement notes, total future payment counts or an eventual closure label is contaminated by future knowledge. Such leakage can produce impressive holdout accuracy even with perfect row-level separation.
Leakage also occurs when near-duplicate contracts, shared properties, catastrophe events or corporate groups are split across training and test data. An embedding may memorize building façade or customer-specific prose across records. In image studies, repeated imagery from the same location or date must be grouped so that the evaluation measures generalization to unseen risks rather than recognition of familiar sites. In claim-narrative studies, boilerplate passages created by the same adjuster or organization can reveal operational identity; cross-adjuster and cross-system tests clarify what has been learned.
A defensible temporal evaluation uses an origin time t, selects training observations available by t, fits encoders and preprocessing only on that training period, and evaluates subsequent contracts or claims on a later period with an outcome-observation horizon appropriate to the product. Cross-validation folds must be constructed before feature selection and hyperparameter tuning. Cawley and Talbot's analysis of selection-induced overfitting (2010) demonstrates why tuning and reporting on the same validation observations can bias performance estimates. Nested or otherwise properly separated evaluation is particularly important when hundreds of encoder architectures, embedding dimensions and pretraining losses are compared.
A practical leakage audit therefore has five layers: a timestamp audit for raw fields, a provenance audit for derived labels and pretrained models, an entity-overlap audit for repeated insured units, a preprocessing audit across the temporal split, and an outcome-maturity audit for the target. These audits should be recorded as substantive evidence of test validity, alongside predictive scores. A technically successful training run supplies none of these guarantees by itself.
Evaluation of Representation Quality
Representations are sometimes evaluated using a convenient proxy objective: reconstruction loss, contrastive alignment, masked-token prediction or cluster separation. These metrics measure how the encoder performs against its training criterion. Insurance usefulness must be established separately against the target exposure, expected loss, relevant claim distribution and operational decisions. Reconstruction of policy descriptions is not proof of accurate insurance pricing; visual similarity of buildings is not proof of similar expected losses; closeness between customer embeddings is not proof that the customers possess comparable insurable risks.
The central empirical design is a controlled comparison in which the unit of observation, training and evaluation periods, claims definition, usable features and computational budget are held fixed. At minimum, the study should report an actuarial baseline such as a Poisson or negative-binomial generalized linear model with exposure offset; an established tree-based model with comparable feature access; a conventional neural model trained from scratch; and the proposed representation-learning configuration. For an additional modality, a baseline using prespecified manual features isolates the gain from adding data, while a pretrained frozen-encoder baseline isolates the gain from task-specific fine-tuning.
Comparative research on general tabular data cautions against presuming architectural superiority. Grinsztajn, Oyallon and Varoquaux (2022) provide empirical comparisons of tree-based and deep models, while Gorishniy and colleagues (2021) establish strong neural tabular baselines through revisiting deep-learning models for tabular data. Both are methodological comparators; neither provides an insurance-wide ranking across lines, jurisdictions and data modalities. The evidence supports experimentation with strong baselines and careful tuning, not a universal claim that neural representations dominate generalized linear models or boosted trees.
For a claim-count model, negative log-likelihood and deviance assess fidelity to count probabilities; exposure-weighted calibration by decile and subgroup shows whether predicted annualized frequency matches observed experience. For severity, proper distributional scoring, mean accuracy and tail-conditional checks are relevant, with the understanding that rare events contribute high sampling variability. For binary claim occurrence, the Brier score is a proper score for probabilities, while area under the receiver-operating-characteristic curve measures discrimination and is invariant under monotone score transformations. The prediction-model evaluation framework of Steyerberg et al. (2010) distinguishes overall performance, discrimination, calibration and clinical utility, providing transferable evaluation principles while its specific clinical decision setting is distinct from insurance. A higher AUC may coexist with defective monetary calibration. For longitudinal targets, evaluation should include the timing and censoring treatment as part of the score definition.
Embedding dimension, data requirements and operational cost belong in the comparison. A latent vector with thousands of coordinates may increase inference latency, storage requirements and audit complexity; a smaller embedding may be adequate for a given line of business. The analyst should provide a performance–complexity curve rather than reporting only the best-fitting architecture. Model parameters, storage of source images or texts, and permitted human review time define different practical constraints and should not be collapsed into a single abstract accuracy score.
Ablations reveal which features produce the result. Separate tests can remove the pretrained initialization, the contrastive loss, the historical sequence, the textual modality, the missingness mask and the supervised risk objective. Results should be presented with confidence intervals or uncertainty estimates based on an appropriate unit of replication, such as policyholder, property, claim cluster, underwriting year or catastrophe event. Bootstrapping individual claim rows as though they were independent can severely understate uncertainty when claims share the same weather event, policyholder or insured structure.
Representational stability can be inspected by perturbation tests, resampling and subgroup diagnostics, but visualizations of an embedding space should be labeled exploratory. Two-dimensional plots produced by dimensionality reduction are influenced by their own optimization and display parameters. Apparent group separation on such a plot is an observation about that projection; it is not automatically an economically meaningful segmentation. The analytical question remains whether the representation improves the probabilistic target on held-out portfolios without creating unjustified use of sensitive or irrelevant variables.
Calibration, Uncertainty, and Tail-Relevant Validation
An insurance decision often depends on monetary expectation, distributional spread and tail exposure. Consequently, the distinction between a ranked score and a calibrated loss distribution is foundational. If a model outputs a claim probability p_i, calibration asks whether observed claim rates among exposures with similar predicted probabilities approximate p_i. If a model predicts annual claim frequency λ_i, calibration asks whether observed claims per exposure align with λ_i. If a model outputs conditional loss quantiles, calibration asks whether the stated coverage probabilities are realized within appropriate risk strata and future cohorts. Different prediction targets require different calibration studies.
Guo and colleagues' modern neural network calibration study (2017) showed miscalibration in several neural classification settings and investigated post-hoc adjustment. Minderer and collaborators' reassessment of neural calibration (2021) demonstrates that model families and evaluation conditions affect the conclusions. These sources motivate empirical probability checks; neither establishes that one post-hoc technique is optimal for all insurance frequency, severity and censored-outcome targets. Calibration procedures must be fitted on a calibration partition separate from final evaluation and should be revisited after material distribution shift.
Insurance portfolios may have thousands of ordinary claims and only a handful of very large losses. Learning from high-frequency ordinary losses can yield reliable mean estimates yet insufficient information about extreme-loss mechanisms. A representation that accurately predicts roofs and construction materials may help a property frequency model while contributing little to tail severity governed by concentrated catastrophe exposure. The literature on images, text and telematics therefore needs an explicit separation between ordinary claim performance and catastrophe, litigation or large-loss tails.
For portfolio loss S=Σ_i L_i, its variance includes both the sum of individual variances and twice the sum of cross-contract covariances: Var(S)=Σ_i Var(L_i)+2Σ_{i<j}Cov(L_i,L_j). A model producing excellent marginal expected losses can still be a poor model of aggregate risk if it ignores common weather, legal, economic or infrastructure shocks. A shared representation may encode observed common exposures, but dependence estimation requires a separate structure and external validation. In a portfolio of buildings exposed to a storm, visually similar features can be useful covariates while the geographic event process remains the dominant source of aggregation.
Parameter uncertainty, stochastic outcome variation and distribution shift should be reported as distinct uncertainties. Repeated training fits measure part of algorithmic variability. A proper predictive distribution measures modeled random loss variability. A comparison across time or territory evaluates transport and instability. Their combined uncertainty cannot be recovered by simply increasing an embedding dimension or training for more epochs. Stress tests should state what changes under the stress: input data, covariate prevalences, claims development, coverage wording, inflation, event hazard or the conditional loss mechanism.
Decision relevance can be formalized by a loss function. For example, under a squared-error objective the conditional mean E[L|X] is optimal, while under an asymmetric loss the appropriate conditional quantile may matter more. A fraud investigation model has an intervention and capacity constraint; a reserving model has a valuation and accounting context; a pricing model must align predicted pure premium with expense, capital and contractual rules. The representation must be evaluated as part of that decision pipeline instead of being granted operational utility from a proxy metric alone.
Interpretability, Fairness, and Legal Context
A latent coordinate is a mathematical feature, and its numerical value has no invariant substantive label without further assumptions. Explanations should therefore distinguish model sensitivity, semantic interpretation, and a causal claim. Sensitivity describes how the prediction changes when an input or embedding coordinate changes under a specified procedure. Semantic interpretation attaches an evidence-supported label to features or patterns. A causal claim requires an intervention-relevant identification design. Moving from the first to the third without supporting assumptions is a category error.
Local attribution methods can identify which measured inputs matter to a fitted model under their formal definition. Lundberg and Lee's unified feature attribution framework (2017) provides a prominent method, but attribution depends on the background distribution and treatment of feature dependence. If customer location, home value, construction quality and historical losses are correlated, reallocating contributions among them does not identify which one causes loss. Attribution over an uninterpreted latent vector often has limited value for a human reviewer; explanations should ideally return to admissible original risk attributes and their documented measurement processes.
Unsupervised learning does not certify semantic disentanglement. Locatello and coauthors' identifiability analysis of disentangled representations (2019) identifies a fundamental absence of unique disentanglement under broad unsupervised assumptions. Insurance researchers cannot assert that an embedding axis specifically represents building hazard, driving skill or claimant honesty merely because the axis correlates with that quantity in one sample. Those interpretations require labeled checks, interventions where appropriate and stability across contexts.
Fairness assessment begins with a legal and operational inventory of the actual decision. The group definition, protected interests, permitted rating variables, claims process and equality criterion depend on jurisdiction and product. An embedding can reproduce correlations between supposedly neutral images, neighborhoods, occupations, health events or language use and protected characteristics. Suppressing an explicit protected field is insufficient to determine whether the representation functions as a proxy. The proper question is whether the resulting use satisfies the relevant insurance rules, contractual and risk-based criteria, and a documented proportionality or fairness assessment.
In the United States, the NAIC's Artificial Intelligence insurance topic materials describe governance issues and the insurance-specific model bulletin adopted by the NAIC in December 2023. Adoption and enforceability depend on action by individual state insurance regulators; it is inaccurate to treat an NAIC model as one universal federal statute. The NIST AI Risk Management Framework (2023) is a voluntary risk-management framework and supplies an organizing vocabulary for governance, mapping, measurement and management. Its voluntary status and scope should remain explicit.
In the European Union, Regulation (EU) 2024/1689 establishes the applicable legal framework, while the AI Act high-risk classification identifies qualifying use in relation to natural persons, subject to the regulation's precise provisions and exemptions. This does not classify every use of machine learning by an insurer as high risk. The EIOPA Opinion on AI governance (2025) addresses governance within the insurance framework for AI systems outside the AI Act's high-risk category and is an interpretive supervisory opinion, not a single globally binding licensing rule. For personal data, the EU General Data Protection Regulation imposes applicable legal requirements on processing, transparency and individual rights, with consequences dependent on the specific processing and jurisdictional facts. This analysis treats primary regulatory texts as jurisdiction-specific evidence rather than a universal checklist.
Practical explanation records should include the target, measurement dates, feature provenance, training population, permitted use, performance stratification, limits of attribution, human review procedures and a route for contesting consequential automated outcomes where the law requires it. A system that can reproduce its mathematical prediction yet cannot identify the source text version, photograph acquisition date or coverage regime is insufficiently traceable for serious insurance use. Fairness and auditability must be tested at the level of the full decision process, including underwriting exclusions and claims handling, rather than at the latent encoder alone.
Robustness, Governance, and Reproducibility
Insurance representation systems are vulnerable to data errors that seldom appear in benchmark train/test protocols. Images may be outdated or altered; claims notes may contain abbreviations, copied boilerplate or unverified allegations; vehicle telematics may be intermittently transmitted; administrative category definitions may change following a merger. Robustness testing should be based on plausible observation failures, not generic perturbations selected only because they are convenient in a machine-learning library. For each input modality, the research record should specify acquisition conditions, expected error mechanisms, version history and permitted retention.
An embedding deployment is a chain of dependencies. Data extraction produces inputs; preprocessing maps values into an accepted schema; a pretrained or proprietary encoder produces latent features; an actuarial head converts those features into predictions; post-processing and business rules transform predictions into decisions. Changing any component can alter effective model behavior even when the headline encoder architecture is unchanged. Version every stage and retain an exact mapping between prediction date and the components operative at that time. For an external pretrained model, preserve model identity, documented license, version, architecture and known pretraining limits when available.
Training governance begins with the declared statistical target and intended portfolio. Evaluate a fixed baseline, predefine the test periods and acceptance criteria, and record every exclusion and recoding. Technical reproducibility includes deterministic preprocessing specifications, declared random seeds, code and dependency versions, training-control settings and stable cryptographic fingerprints for the manuscript and evidence record. Scientific reproducibility additionally requires that the sample construction, research estimand, validation population and evidence limitations be intelligible to a qualified independent investigator. A reproducible computation of an invalid estimand remains scientifically invalid.
Monitoring should measure input schema failures, sample composition shifts, unexpected absence of modalities, score distribution changes and outcome-linked performance once claims mature. A shift alarm is evidence of a potential issue and should trigger diagnosis; it does not establish the cause of the shift. Particular attention is required where deployment changes participation: telematics discounts may alter enrollment; an inspection policy may alter which properties have images; and fraud referrals may change subsequent claims data. The resulting feedback can transform the dataset used for later training.
The choice of a supervised, contrastive, masked-token or autoencoding objective should be justified with respect to useful signal, sample selection and stable deployment data. No source reviewed here establishes a universal representation objective for all insurance lines. In an actuarial setting, a relatively simple tabular model can be more governable and equally accurate; elsewhere a multimodal encoder can recover a measured signal unavailable to that model. Evidence-based governance permits both outcomes and requires the choice to be demonstrated on the relevant portfolio.
Worked Analytical Illustrations
The following examples are original transparent mathematical illustrations. They do not describe empirical results, product prices or validated risk predictions.
Consider two hypothetical insured groups A and B, with 1,200 and 800 policy-years of exposure. If the true expected claim frequencies are respectively 0.08 and 0.10 claims per policy-year, expected claim counts are 96 and 80. With a shared illustrative conditional mean severity of $5,000, corresponding expected aggregate claims are $480,000 and $400,000. The combined portfolio has 2,000 policy-years and $880,000 in expected claims, or $440 per policy-year. If a representation-driven frequency model wrongly assigns group A the frequency 0.10, it adds 24 expected claims, increases estimated claims by $120,000 and changes combined expected claims to $1,000,000. This illustrates why a prediction model must preserve exposure arithmetic and segment-level calibration even when a latent embedding improves ranking. The assumed frequencies and severities are illustrative inputs, not estimates from insurance data.
For a second example, let the risk outcome Y satisfy E[Y|X]=m(X), and let Z=f(X). The best squared-error predictor using only the representation is E[Y|Z]. The law of total variance implies E[(Y−E[Y|Z])²]=E[Var(Y|Z)] and E[(Y−E[Y|X])²]=E[Var(Y|X)]. Since Z is determined by X, conditioning on the full observed X cannot have greater optimal squared-error risk than conditioning on Z: E[Var(Y|Z)]≥E[Var(Y|X)]. Thus, deterministic compression cannot increase the population-level information available to the ideal predictor. Yet a learned representation can improve finite-sample predictions because it supplies regularization, pooling, pretrained structure or a more manageable function class. The apparent paradox dissolves when statistical estimation error is distinguished from information content. This is an analytical consequence of conditional expectation, not a new doctrine or empirical superiority claim.
A third example concerns missing imagery. Suppose inspection images exist for 30% of insured buildings and those imaged have an observed claim frequency of 0.06, while uninspected buildings have a frequency of 0.11. The policy-year-weighted overall frequency would be 0.30×0.06+0.70×0.11=0.095, given equal exposure units. A visual model evaluated only on imaged buildings can achieve excellent calibration at 0.06 and nevertheless fail to address the 70% without images. For an operational system, the missing-modality route and the assessment of the full portfolio are indispensable. The numbers simply demonstrate the sampling distinction.
Finally, consider a latent representation Z=WX for an invertible square matrix W, with a linear predictor βᵀZ. For any invertible matrix A, define Z'=AZ and β'=A^{−ᵀ}β. Then β'ᵀZ'=βᵀZ. Identical predictions can therefore arise from different latent coordinate systems. Even in this elementary setting, interpreting one coordinate as a uniquely identifiable risk factor is unjustified without additional restrictions. This illustration complements the broader representation-identifiability literature, including Locatello et al. (2019).
Comparative Evidence and Research Agenda
The evidence surveyed supports three qualified conclusions about current methods. First, representation learning is mathematically well established as a family of feature-learning objectives, including denoising, language modeling, contrastive alignment and attention-based tabular encoding, as detailed by Bengio et al. (2013), Vincent et al. (2008), Devlin et al. (2019) and Chen et al. (2020). These mechanisms give researchers a systematic language for learning intermediate predictive variables from observed data.
Second, published insurance applications show that particular representations are feasible in identified datasets: high-cardinality categorical risk records, property images, and telematics series are studied in actuarial research, including Avanzi et al. (2024), Blier-Wong et al. (2024) and Duval et al. (2024). Each paper's investigated outcome, available modality, loss definition and sampling design constrain the conclusion. An empirical case study does not by itself establish cross-market dominance of a representation family.
Third, reliability depends on data and institutional context. The out-of-distribution methods literature, calibration research and regulatory materials provide independent reasons to test temporal transfer, outcome-specific probability calibration and actual decision governance. Their relevance to insurance is a methodologically argued synthesis, not a claim that their general-domain benchmarks have measured insurance results. The Peiris et al. (2024) data-integration study is useful for thinking about combined traditional and telematics information; its synthetic-data setting limits direct market generalization.
A cumulative research program should prioritize preregistered or otherwise prespecified out-of-time evaluations on privacy-appropriate real portfolios, documentation of missing modalities, honest comparisons against strong actuarial baselines, standardized claim maturity and exposure normalization, and subgroup-calibrated loss estimates. Multi-institutional replications should distinguish product design and jurisdiction rather than pooling incompatible outcomes under a common label. For visual and linguistic models, researchers should document the age and provenance of observations and the relationship between features and insurable hazards. For transfer methods, controlled studies should compare training-data budgets and show when pretrained embeddings help or harm.
Additional questions warrant study. Which representation objectives yield measurable gains with small labeled samples? How stable are learned claim-narrative features across new adjuster practices and multilingual markets? How much apparent gain depends on recorded inspection selection? Can a shared representation improve ordinary claims while underrepresenting catastrophe exposure? How do fairness safeguards and confidentiality controls alter the effective training distribution? These are proposed empirical questions. They are not announcements of newly founded theories, named principles or established causal findings.
Conclusion
Deep representation learning offers a general set of methods for converting structured risk records, text, images and longitudinal observations into features that can enter insurance estimation. Its central scientific value is the possibility of learning useful feature transformations from relevant data and, where justified, transferring information across tasks or data modalities. What a learned coordinate represents remains conditional on the training objective, observed variables, sample-selection mechanism and identification assumptions. Its predictive usefulness is specific to an outcome and a population.
The insurance application establishes a disciplined sequence: define a valid contractual and actuarial target; map exposures, report delays, missing inputs and permissible information; learn a representation from properly partitioned data; compare it with strong actuarial and machine-learning alternatives; and verify calibration, robustness and uncertainty on held-out portfolios and periods. The representation's assessment must include the insurer's decision process, geographical and temporal deployment, and applicable legal obligations. An improvement in a general pretraining loss cannot substitute for measured performance on claim frequencies, severities, ultimate losses or another properly defined insurance outcome.
The primary literature provides demonstrated methodological foundations and bounded empirical insurance examples. Their synthesis supports targeted adoption when a representation supplies incremental information, appropriate calibration, stable generalization and traceable governance. It also supports using simpler methods when their evidence is stronger. A successful insurance representation is established by measured actuarial usefulness under disclosed assumptions, alongside replicable source and data provenance. The outstanding task for cumulative research is to determine those conditions across independently studied products, markets and time periods.
References
Arik, S. Ö., & Pfister, T. (2021). TabNet: Attentive interpretable tabular learning. Proceedings of the AAAI Conference on Artificial Intelligence, 35(8), 6679–6687. https://ojs.aaai.org/index.php/AAAI/article/view/16826
Avanzi, B., Taylor, G., Wang, M., & Wong, B. (2024). Machine learning with high-cardinality categorical features in actuarial applications. ASTIN Bulletin, 54(2), 213–238. DOI:10.1017/asb.2024.7. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/machine-learning-with-highcardinality-categorical-features-in-actuarial-applications/C910580D669A903DA6CC00AE6CDCB4DE
Bengio, Y., Courville, A., & Vincent, P. (2013). Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8), 1798–1828. DOI:10.1109/TPAMI.2013.50. https://pubmed.ncbi.nlm.nih.gov/23787338/
Blier-Wong, C., Lamontagne, L., & Marceau, E. (2024). A representation-learning approach for insurance pricing with images. ASTIN Bulletin, 54(2), 280–309. DOI:10.1017/asb.2024.9. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/representationlearning-approach-for-insurance-pricing-with-images/E442676A2F597A9DBBDFC9FB986C3070
Cawley, G. C., & Talbot, N. L. C. (2010). On over-fitting in model selection and subsequent selection bias in performance evaluation. Journal of Machine Learning Research, 11, 2079–2107. https://jmlr.org/papers/v11/cawley10a.html
Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. (2020). A simple framework for contrastive learning of visual representations. Proceedings of Machine Learning Research, 119, 1597–1607. https://proceedings.mlr.press/v119/chen20j.html
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of NAACL-HLT 2019, 4171–4186. DOI:10.18653/v1/N19-1423. https://aclanthology.org/N19-1423/
Duval, F., Boucher, J.-P., & Pigeon, M. (2024). Telematics combined actuarial neural networks for cross-sectional and longitudinal claim count data. ASTIN Bulletin, 54(2), 239–262. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/telematics-combined-actuarial-neural-networks-for-crosssectional-and-longitudinal-claim-count-data/B6C01BF508F64D4C7A804A628F3D03E0
European Commission. (n.d.; accessed November 15, 2026). AI Act Service Desk: Essential services, Annex III point 5(c). https://ai-act-service-desk.ec.europa.eu/en/essential-services
European Insurance and Occupational Pensions Authority. (2025, August 6). Opinion on artificial intelligence governance and risk management (EIOPA-BoS-25-360). https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en
European Parliament and Council of the European Union. (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). Official Journal of the European Union, L 119, 1–88. https://eur-lex.europa.eu/legal-content/EN/TXT/?qid=1678624835077&uri=CELEX%3A32016R0679
European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union, L, 2024/1689. https://eur-lex.europa.eu/eli/reg/2024/1689
Gorishniy, Y., Rubachev, I., Khrulkov, V., & Babenko, A. (2021). Revisiting deep learning models for tabular data. Advances in Neural Information Processing Systems, 34. https://proceedings.neurips.cc/paper_files/paper/2021/hash/9d86d83f925f2149e9edb0ac3b49229c-Abstract.html
Grinsztajn, L., Oyallon, E., & Varoquaux, G. (2022). Why do tree-based models still outperform deep learning on typical tabular data? Advances in Neural Information Processing Systems, 35, Datasets and Benchmarks Track. https://proceedings.nips.cc/paper_files/paper/2022/hash/0378c7692da36807bdec87ab043cdadc-Abstract-Datasets_and_Benchmarks.html
Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On calibration of modern neural networks. Proceedings of Machine Learning Research, 70, 1321–1330. https://proceedings.mlr.press/v70/guo17a.html
Koh, P. W., Sagawa, S., Marklund, H., Xie, S. M., Zhang, M., et al. (2021). WILDS: A benchmark of in-the-wild distribution shifts. Proceedings of Machine Learning Research, 139, 5637–5664. https://proceedings.mlr.press/v139/koh21a.html
Locatello, F., Bauer, S., Lucic, M., Rätsch, G., Gelly, S., Schölkopf, B., & Bachem, O. (2019). Challenging common assumptions in the unsupervised learning of disentangled representations. Proceedings of Machine Learning Research, 97, 4114–4124. https://proceedings.mlr.press/v97/locatello19a.html
Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30. https://proceedings.neurips.cc/paper_files/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html
Minderer, M., Djolonga, J., Romijnders, R., Hubis, F., Zhai, X., Houlsby, N., Tran, D., & Lucic, M. (2021). Revisiting the calibration of modern neural networks. Advances in Neural Information Processing Systems, 34. https://proceedings.neurips.cc/paper/2021/hash/8420d359404024567b5aefda1231af24-Abstract.html
National Association of Insurance Commissioners. (2023; current topic page accessed November 15, 2026). Model Bulletin: Use of Artificial Intelligence Systems by Insurers (adopted December 4, 2023); AI topic and state adoption guidance. https://content.naic.org/insurance-topics/artificial-intelligence
National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0) (NIST AI 100-1; E. Tabassi, author). DOI:10.6028/NIST.AI.100-1. https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10
Peiris, H., Jeong, H., Kim, J.-K., & Lee, H. (2024). Integration of traditional and telematics data for efficient insurance claims prediction. ASTIN Bulletin, 54(2), 263–279. DOI:10.1017/asb.2024.6. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/integration-of-traditional-and-telematics-data-for-efficient-insurance-claims-prediction/2D72787C1DE75B6DF0FECC90E6414B40
Quiñonero-Candela, J., Sugiyama, M., Schwaighofer, A., & Lawrence, N. D. (Eds.). (2009). Dataset shift in machine learning. MIT Press. DOI:10.7551/mitpress/9780262170055.001.0001. https://academic.oup.com/mit-press-scholarship-online/book/13447
Steyerberg, E. W., Vickers, A. J., Cook, N. R., Gerds, T., Gonen, M., Obuchowski, N., Pencina, M. J., & Kattan, M. W. (2010). Assessing the performance of prediction models: A framework for some traditional and novel measures. Epidemiology, 21(1), 128–138. DOI:10.1097/EDE.0b013e3181c30fb2. https://pmc.ncbi.nlm.nih.gov/articles/PMC3575184/
Vincent, P., Larochelle, H., Bengio, Y., & Manzagol, P.-A. (2008). Extracting and composing robust features with denoising autoencoders. Proceedings of the 25th International Conference on Machine Learning, 1096–1103. https://doi.org/10.1145/1390156.1390294
Wüthrich, M. V., & Merz, M. (2019). Editorial: Yes, we CANN! ASTIN Bulletin, 49(1), 1–3. DOI:10.1017/asb.2018.42. https://www.cambridge.org/core/journals/astin-bulletin-journal-of-the-iaa/article/editorial-yes-we-cann/66E8BEC373B5CCEF3BF3303D442D6B75